TypeSafe AI’s Jev returns typed decisions with a calibrated confidence score, fast and cheap enough to sit inside of virtually every step of a production workflow. We’ve spent its first week testing use-cases, and it changes what an AI system can afford to decide by design.
Jev, released last week by TypeSafe AI, is arguably the most interesting release of the year for anyone building production AI systems. It takes an entirely different approach from the LLMs that have defined our collective understanding of AI use-cases over the past three years: decisions instead of language, milliseconds vs. seconds, and fractions of a cent instead of cents to do its job.
TypeSafe calls Jev a “System One” model, after Daniel Kahneman’s term for fast, intuitive judgment. Software hands it unstructured input and a question defined in advance, and Jev returns a typed decision with a probability attached, rather than a paragraph of text for a person to read. TypeSafe reports responses in 70 to 500 milliseconds at $0.042 per million input tokens, with output free.
Our AI Innovation teams started putting Jev through its paces as soon as the early release opened, and a week in we’ve now run it across semantic classification, document routing and answer scoring. It does what TypeSafe says it does: fast, structured decisions at a cost and latency that no LLM was going to reach. That’s the pitch, and in our testing the pitch holds up.
The more important point here is actually architectural. Most of what a business workflow does is deciding rather than thinking: which queue, which reviewer, which next step. LLMs have steered us a bit wrong on this point. You don’t need creative intelligence just to put intelligence to work in automation.
Below we cover why those decisions have been expensive to automate, what changes when they suddenly cost effectively nothing, why the confidence score matters more than the speed, and where we’re already putting Jev to work.
Until now, teams had to choose between fast and flexible
Before Jev, a team automating a routine decision could have fast and cheap, or it could have flexible, but never both (and often without a confidence level that satisfied at scale). For that reason, teams might have scoped their ambition down to fit whichever compromise they picked, or may have limited their own imagination on automation use-cases overall.
Consider how Jev affects our options:
- The trained classifier: A small purpose-built model is fast, cheap and accurate, but its categories are fixed at training time, so every policy change requires fresh labeled examples and a retraining cycle. In a routing scenario we tested where the rule changed mid-stream, the trained classifier kept confidently (but wrongly) applying the old policy, while Jev applied the new rule on the next request because the categories live in the question (think, prompt design) rather than in the model’s weights.
- The general-purpose LLM: A frontier model is flexible and needs no training data, but it is priced and paced for working with a conversation rather than a pipeline, at seconds per decision and 15 to 185x the cost of Jev. At a certain volume, that math forces compromises: run the process nightly, sample a slice of the transactions, or reserve the model for escalations. The capability existed, sure, but pre-Jev it was hard to justify intelligent automation for many use-cases, at real scale.
When a decision costs almost nothing, a workflow can afford many more of them
When routine decisions are handled by a model built for them, the process can run on every transaction instead of batching or working with a sample, with the expensive reasoning of an LLM is reserved for the cases that are hard.
This is a design implication we have not seen discussed much so far. Work that teams handed to an LLM because there was no better option can now move to a model built for it. This lets us start imagining new possibilities, including things like:
- Automated testing: classifying test results and defects and routing each one to the exact right owner, efficiently and real-time
- Adaptive UI: deciding what an interface should show based on what a user is doing, fast enough to happen between clicks, as if predictively
- Visual detection: flagging and sorting images at a volume where per-call LLM pricing would be prohibitive
- Fraud and security screening: scoring every event instead of a sample
The lower AI bill is the smaller benefit. The larger one is that more work moves through without a human touch, under a control that risk and compliance teams will approve, while the organization’s best people spend their time on the judgment calls that matter.
Calibrated confidence is what separates Jev from earlier classifiers
What separates Jev from earlier classifier approaches is more than speed: every decision comes with a probability attached, so a system receives a confidence number along with the verdict. That might sound minor, and it’s definitely not. Think of it like this: a system that can act on “91 percent sure” gets designed differently from one that only sees yes or no.
A yes-or-no classifier sends every output down the same path, while a scored decision lets the system choose the path based on how sure it is, and that sureness is very accurate. With a calibrated score, a workflow can automate the clear-cut cases outright, send a middle band to quick review, and route only the uncertain cases to a person.
Note: We’d still recalibrate against client data before trusting any general threshold. TypeSafe trains Jev specifically for calibrated confidence, but a 0.9 on a vendor benchmark and a 0.9 on one of our client’s claims queues are different measurements.
The same property changes what agentic systems can be built on.
The pattern we find most promising pairs the two kinds of model inside an agentic workflow: the LLM proposes the schema, and the decision layer populates and maintains it. Generation is System 2 work and Population is System 1 work, which gives the two models a clean division of the jobs an agentic workflow has to do. Left brain right brain, if you will.
Probabilistic knowledge graphs are a well-established research idea that has mostly stayed in research, because computing an honest confidence for every edge at scale was uneconomic. But now that every assertion in an ontology can include a cheaply computed, honestly calibrated confidence weight, three capabilities become practical:
- Preferential traversal: Agents can traverse the graph along high-confidence paths, so task decomposition follows the parts of the domain model the organization actually trusts.
- Routing by uncertainty: The system can auto-accept the confident tails and send only the ambiguous middle band to a person or a frontier model, which means expensive tokens are spent exclusively on genuine ambiguity.
- Continuous re-scoring: The graph can be re-scored continuously, so a team can watch the confidence distribution over a subgraph fall when the code underneath it changes and treat that drop as an early warning.
At 3Pillar, we’re now exploring Jev within key delivery systems
The best use we have found so far is embedded rather than standalone: high-volume, low-cost, high-speed classification inside the applications we build and ship for clients. Jev is currently at work in three places:
- Next best action in HelixAI: Jev identifies incoming emails and information so that Helix can choose the right next step for each one.
- Compliance checks: Jev makes fast, scored determinations on content before it moves downstream.
- Model routing: Our AI Innovation team has been prototyping a router that lets Jev decide which model should handle a given request instead of routing on static rules, and the early results from those prototypes are good.
The application we are most excited about is using Jev as a step decider and guardrail inside intelligent workflow automation, where an agent checks a fast, scored decision at each step and escalates when confidence drops.
Model availability and enterprise hosting are still unresolved, and we are not going to pretend otherwise. Jev is in early access behind a waitlist, TypeSafe has not yet offered self-hosting, and TypeSafe itself notes that current pricing may be subsidized.
For clients with strict data residency requirements or regulated workloads, those questions have to be answered before production. We also expect a wave of specialized local alternatives over the next few months, which would address much of the hosting concern.
The advantage is moving from model access to process knowledge
This looks to us like the first real fork in the model market since LLMs took over. Jev is a legitimately different kind of model from an LLM, built to decide rather than write. It does not replace the creativity and full reasoning an LLM provides; it complements that reasoning and frees it for the work that actually needs it.
The last three years advantaged whoever had the best access to the biggest model. A decision layer advantages whoever understands a client’s process well enough to break a decision into its parts, and who is disciplined enough to recalibrate confidence before trusting it. That is an architecture and domain-knowledge game, and it is the one we would rather be playing.
Explore it with us
If you run an AI workflow that performs in a pilot but becomes expensive at volume, we’d like to look at it with you. Our Helix Forward team can map which steps are really best understood as decisions, test Jev against your data, and show where confidence thresholds would let you automate more. Let’s talk about it.
See how we build agentic systems with HelixAI and AI-Powered Operations, or contact 3Pillar to talk with our team.
